The person consuming the most AI tokens inside one of the world’s most advanced AI companies isn’t an engineer, a researcher, or a product manager. It’s the head of tax.
Anthropic CFO Krishna Rao made the reveal during a May 13 appearance on the Invest Like the Best podcast, noting that senior finance professionals have become the heaviest internal users of Claude, the company’s flagship AI model.
The finance team’s AI playbook
Rao described a finance organization that has gone far beyond casual AI experimentation. The team maintains between 70 and 150 custom Claude “skills,” which are essentially purpose-built workflows and agents stored in shared repositories similar to GitHub.
Statutory financial statements, the kind of documents that traditionally require painstaking human assembly, are now predominantly generated by Claude with human review layered on top. Weekly revenue and compute utilization reports that used to consume several hours now take roughly 30 minutes.
Rao also mentioned that Anthropic tracks internal token usage through leaderboards. These are purely informational and aren’t tied to compensation or performance reviews.
Why the tax team, of all people
The fact that tax professionals emerged as the top token consumers tells a deeper story about where AI delivers the most value in enterprise settings. Tax work is notoriously complex, document-heavy, and governed by intricate regulatory frameworks that shift constantly.
A tax expert knows exactly what questions to ask and can immediately evaluate whether an AI-generated answer is correct. The finance professionals who already understood their domain deeply were the ones best positioned to build those 70 to 150 custom skills and integrate them into daily workflows.
What this means for investors and the broader market
The productivity gains Rao described—hours-long reports compressed to 30 minutes, automated financial statement generation—are measurable efficiency improvements. If Anthropic’s finance team can maintain up to 150 custom AI workflows, the scale of what larger finance organizations could build is significant.
As finance teams automate critical processes like statutory reporting and tax analysis, the stakes around AI accuracy get significantly higher. A hallucinated number in a statutory financial statement is a potential regulatory problem. The “human review as needed” framework Rao described will need to remain genuinely rigorous.
Rao also pointed to an expectation that newer, more capable Claude models will drive even higher token consumption. As the models get better, the returns on each query improve, which encourages more usage—a dynamic that directly links product improvements to revenue growth from existing users.